[
https://issues.apache.org/jira/browse/HADOOP-3810?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Hairong Kuang updated HADOOP-3810:
----------------------------------
Attachment: globalLock.patch
This patch aims to improve NN responsiveness when a cluster is near full. It
identified one cause of the problem is that ReplicationMonitor holds the global
lock while replicating blocks in one iteration. So this patch made the
following change:
1. does not allow ReplicationMonitor to hold the fsnamesystem global lock while
replicating block in one iteration. Previously the logic is like:
{code}
synchronized computeReplicationWork( int blocksToProcess) {
for (int i=0; i<blocksToProcess; i++) {
select one block from under-replicated queue;
Compute replication work for the block;
}
}
{code}
Now it is changed to be:
{code}
computeReplicationWork(int blocksToProcess) {
(synchronized) select blocksToProcess under-replicated blocks;
for each selected block {
(synchronized)compute replication work for the block;
}
}
{code}
2. While computing replication work for a block, releases the global lock while
choosing targets because this is the most computation intensive part.
> NameNode seems unstable on a cluster with little space left
> -----------------------------------------------------------
>
> Key: HADOOP-3810
> URL: https://issues.apache.org/jira/browse/HADOOP-3810
> Project: Hadoop Core
> Issue Type: Bug
> Components: dfs
> Affects Versions: 0.17.1
> Reporter: Raghu Angadi
> Assignee: Hairong Kuang
> Fix For: 0.21.0
>
> Attachments: globalLock.patch, simon-namenode.PNG
>
>
> NameNode seems not very responsive and unstable when the cluster has very
> little space left. The clients timeout. The main problem is that it is not
> clear to the user what is going on. Once I have more details about a NameNode
> that was in this state, I will fill in here.
> If there is not enough space left on a cluster, it is ok for clients to
> receive something like "DiskOutOfSpace" exception.
> Right now it looks like NameNode tries too hard find a node with any space
> left and ends up being slow to respond to clients. If the CPU taken by
> chooseTarger() is the main cause, there are two possible fixes :
> # chooseTarget() iterates and takes quite a bit of CPU for allocating
> datanodes. Usually this not much of a problem. It takes even more cpu when it
> needs to search multiple racks for a datanode. We could probably reduce some
> CPU for these searches. The benefit should be measurable.
> # Once NameNode can not find any datanode that has space on a rack, it could
> mark the rack as "full" and skip searching the rack for next one minute or
> so. This flag gets cleared after a minute or if any new node is added to the
> rack.
> #* Of course, this might not be optimal w.r.t disk space usage.. but only for
> a short duration. Once a cluster is mostly full, the user does expect errors.
> #* On the flip side, this fix does not require extremely CPU optimized
> version of chooseTarget().
> #* I think it is reasonable for NameNode to throw DiskOutOfSpace exception,
> even though it could have found space if it searched much more extensively.
> ---
> edit : minor changes
>
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
